SafePointNodes and raw memory

[Tom Rodriguez]

I was very surprised today to discover that SafePointNodes don't  
create a new raw memory state.  Does anyone know the history of this?   
This breaks any piece of code using raw memory that relies on memory  
dependence for proper evaluation order which is a pretty basic  
assumption of ideal.  The G1 pre barriers are broken at least partly  
for this reason.

tom